Spaces:
Running
on
Zero
Running
on
Zero
Upload 13 files
Browse files- .gitattributes +3 -0
- ACLlama_el_s2s.py +565 -0
- EchoX-Vocoder/checkpoint_last.pt +3 -0
- EchoX-Vocoder/config.json +53 -0
- EchoX-Vocoder/g_00500000 +3 -0
- EchoX-Vocoder/re_config.log +0 -0
- EchoX-Vocoder/spm_1k.model +3 -0
- Echox_copy_stream.py +426 -0
- T2ULlama_CR_online.py +412 -0
- show_case/1.wav +0 -0
- show_case/2.wav +3 -0
- show_case/Translate_de_audio_prompt.wav +3 -0
- text_to_speech.py +326 -0
.gitattributes
CHANGED
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
EchoX-Vocoder/g_00500000 filter=lfs diff=lfs merge=lfs -text
|
37 |
+
show_case/2.wav filter=lfs diff=lfs merge=lfs -text
|
38 |
+
show_case/Translate_de_audio_prompt.wav filter=lfs diff=lfs merge=lfs -text
|
ACLlama_el_s2s.py
ADDED
@@ -0,0 +1,565 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional, Tuple, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
from torch.nn import CrossEntropyLoss, CTCLoss
|
7 |
+
|
8 |
+
|
9 |
+
from transformers import AutoConfig, AutoModelForCausalLM, \
|
10 |
+
LlamaConfig, LlamaModel, LlamaForCausalLM
|
11 |
+
from transformers.trainer_pt_utils import LabelSmoother
|
12 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
|
13 |
+
from transformers import (
|
14 |
+
WhisperProcessor,
|
15 |
+
WhisperModel,
|
16 |
+
)
|
17 |
+
from T2ULlama_CR_online import T2ULlamaForCausalLM
|
18 |
+
|
19 |
+
IGNORE_TOKEN_ID = LabelSmoother.ignore_index
|
20 |
+
|
21 |
+
class ACLlamaConfig(LlamaConfig):
|
22 |
+
model_type = "ACLlama"
|
23 |
+
|
24 |
+
def load_whisper(audio_tower_name, device="cuda"):
|
25 |
+
model = WhisperModel.from_pretrained(
|
26 |
+
audio_tower_name,torch_dtype=torch.float16,low_cpu_mem_usage=True).to(device)
|
27 |
+
model.config.forced_decoder_ids = None
|
28 |
+
return model
|
29 |
+
|
30 |
+
class LookBackModule(nn.Module):
|
31 |
+
def __init__(self, cfg: LlamaConfig):
|
32 |
+
super().__init__()
|
33 |
+
self.encoder_attn = nn.MultiheadAttention(
|
34 |
+
cfg.hidden_size,
|
35 |
+
cfg.num_attention_heads,
|
36 |
+
dropout=0.1,
|
37 |
+
batch_first=True
|
38 |
+
)
|
39 |
+
self.atten_layer_norm = nn.LayerNorm(cfg.hidden_size)
|
40 |
+
|
41 |
+
|
42 |
+
def forward(self, x, wav_feature, bf_shrink_padding_mask):
|
43 |
+
|
44 |
+
residual = x
|
45 |
+
x, _ = self.encoder_attn(
|
46 |
+
query=x,
|
47 |
+
key=wav_feature,
|
48 |
+
value=wav_feature,
|
49 |
+
key_padding_mask=bf_shrink_padding_mask,
|
50 |
+
#attn_mask=padding_mask,
|
51 |
+
)
|
52 |
+
x += residual
|
53 |
+
x = self.atten_layer_norm(x)
|
54 |
+
return x
|
55 |
+
|
56 |
+
class ACLlamaModel(LlamaModel):
|
57 |
+
config_class = ACLlamaConfig
|
58 |
+
|
59 |
+
def __init__(self, config: LlamaConfig):
|
60 |
+
super(ACLlamaModel, self).__init__(config)
|
61 |
+
|
62 |
+
if hasattr(config, "audio_tower"):
|
63 |
+
self.audio_tower = [load_whisper(config.audio_tower)]
|
64 |
+
|
65 |
+
if hasattr(config, "adapter_size"):
|
66 |
+
|
67 |
+
self.mm_projector1 = nn.Linear(config.adapter_size*2 , config.hidden_size)
|
68 |
+
asr_encoder_layer = nn.TransformerEncoderLayer(
|
69 |
+
d_model=config.hidden_size,
|
70 |
+
nhead=config.num_attention_heads,
|
71 |
+
dim_feedforward=config.hidden_size*2,
|
72 |
+
dropout=0.1,
|
73 |
+
norm_first=True
|
74 |
+
)
|
75 |
+
self.lbm = LookBackModule(config)
|
76 |
+
self.out_norm = nn.LayerNorm(config.hidden_size)
|
77 |
+
self.audio_feature_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
78 |
+
self.asr_transformer_encoder = nn.TransformerEncoder(asr_encoder_layer, num_layers=1)
|
79 |
+
self.mask_tensor=(torch.ones([1, 2048])>0)
|
80 |
+
self.length=-1
|
81 |
+
|
82 |
+
def forward(
|
83 |
+
self,
|
84 |
+
input_ids: torch.LongTensor = None,
|
85 |
+
attention_mask: Optional[torch.Tensor] = None,
|
86 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
87 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
88 |
+
use_cache: Optional[bool] = None,
|
89 |
+
output_attentions: Optional[bool] = None,
|
90 |
+
output_hidden_states: Optional[bool] = None,
|
91 |
+
audios: Optional[torch.FloatTensor] = None,
|
92 |
+
return_dict: Optional[bool] = None,
|
93 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
94 |
+
|
95 |
+
# HACK: replace back original embeddings for LLaAA pretraining
|
96 |
+
orig_embeds_params = getattr(self, 'orig_embeds_params', None)
|
97 |
+
|
98 |
+
if inputs_embeds is None:
|
99 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
100 |
+
|
101 |
+
audio_tower = getattr(self, 'audio_tower', None)
|
102 |
+
if audio_tower is not None and (input_ids.shape[1] != 1 or self.training) and audios is not None:
|
103 |
+
audio_tower = audio_tower[0] # HACK: for FSDP
|
104 |
+
audio_list=[]
|
105 |
+
|
106 |
+
audio_config = audio_tower.config
|
107 |
+
for audio in audios:
|
108 |
+
with torch.no_grad():
|
109 |
+
audio_feature = audio_tower.encoder(audio).last_hidden_state
|
110 |
+
|
111 |
+
audio_feature = audio_feature.view(audio_feature.shape[0], audio_feature.shape[1]//2, 2 * audio_feature.shape[2])
|
112 |
+
audio_feature = self.mm_projector1(audio_feature)
|
113 |
+
audio_feature = self.asr_transformer_encoder(audio_feature)
|
114 |
+
audio_feature = self.out_norm(audio_feature)
|
115 |
+
audio_list.append(audio_feature)
|
116 |
+
|
117 |
+
audio_features = torch.stack(audio_list, dim=0)
|
118 |
+
batch = audio_features.shape[0]
|
119 |
+
audio_turn = audio_features.shape[1]
|
120 |
+
audio_features = audio_features.view((batch * audio_turn,)+audio_features.shape[2:])
|
121 |
+
|
122 |
+
predict_logits = self.audio_feature_head(audio_features)
|
123 |
+
|
124 |
+
new_input_embeds = []
|
125 |
+
label_shift = []
|
126 |
+
speech_pos = []
|
127 |
+
label_extend = -1
|
128 |
+
new_input_ids = []
|
129 |
+
tokens = predict_logits.argmax(dim=-1)
|
130 |
+
shrink_mask = tokens.roll(1) != tokens
|
131 |
+
shrink_mask[:,0] = True
|
132 |
+
|
133 |
+
lengths = shrink_mask.long().sum(-1)
|
134 |
+
shrink_2d = audio_features[shrink_mask]
|
135 |
+
#num_patches = audio_features.shape[1]
|
136 |
+
num_patches = audio_config.audio_patch_size
|
137 |
+
l_index=0
|
138 |
+
shrink_features_raw = []
|
139 |
+
for v, audio_feature, mask in zip(lengths, audio_features, ~shrink_mask):
|
140 |
+
shrink_feature = shrink_2d[l_index:l_index+v]
|
141 |
+
shrink_feature = self.lbm(shrink_feature, audio_feature, bf_shrink_padding_mask=mask)
|
142 |
+
shrink_features_raw.append(shrink_feature)
|
143 |
+
l_index += v
|
144 |
+
|
145 |
+
shrink_features = []
|
146 |
+
for i in range(0, len(shrink_features_raw), audio_turn):
|
147 |
+
shrink_features.append(shrink_features_raw[i:i+audio_turn])
|
148 |
+
if self.training:
|
149 |
+
maxn_length = lengths.view(batch,audio_turn).sum(-1).max()
|
150 |
+
label_extend = maxn_length - num_patches * audio_turn
|
151 |
+
old_seq_length = inputs_embeds.shape[1]
|
152 |
+
for cur_input_ids, cur_input_embeds, cur_shrink_features in zip(input_ids, inputs_embeds, shrink_features):
|
153 |
+
pad_ids = torch.full(size=(maxn_length,), fill_value=audio_config.llm_pad_token_id, dtype=torch.long).to(attention_mask.device)
|
154 |
+
pad_embeds = self.embed_tokens(pad_ids)
|
155 |
+
audio_start_token_pos_all = torch.where(cur_input_ids == audio_config.audio_patch_token)[0]
|
156 |
+
#print(cur_input_embeds.shape,cur_input_ids.shape)
|
157 |
+
inner_label_shift = []
|
158 |
+
inner_speech_pos = []
|
159 |
+
for audio_start_token_pos, shrink_feature in reversed(list(zip(audio_start_token_pos_all, cur_shrink_features))): #zip(audio_start_token_pos_all, cur_shrink_features):
|
160 |
+
cur_speech_length = shrink_feature.shape[0]
|
161 |
+
|
162 |
+
cur_input_ids = torch.cat((cur_input_ids[:audio_start_token_pos],
|
163 |
+
cur_input_ids[audio_start_token_pos: audio_start_token_pos+1].repeat(cur_speech_length),
|
164 |
+
cur_input_ids[audio_start_token_pos + num_patches:]), dim=0)
|
165 |
+
cur_input_embeds = torch.cat((
|
166 |
+
cur_input_embeds[:audio_start_token_pos],
|
167 |
+
shrink_feature,
|
168 |
+
cur_input_embeds[audio_start_token_pos + num_patches:]), dim=0)
|
169 |
+
inner_label_shift.insert(0, cur_speech_length - num_patches)
|
170 |
+
inner_speech_pos.insert(0, audio_start_token_pos)
|
171 |
+
|
172 |
+
label_shift = label_shift + inner_label_shift
|
173 |
+
speech_pos = speech_pos + inner_speech_pos
|
174 |
+
|
175 |
+
cur_new_input_embeds = torch.cat((cur_input_embeds, pad_embeds[:old_seq_length + label_extend - cur_input_embeds.shape[0]]),dim=0)
|
176 |
+
cur_new_input_ids = torch.cat((cur_input_ids, pad_ids[:old_seq_length + label_extend - cur_input_ids.shape[0]]),dim=0)
|
177 |
+
new_input_embeds.append(cur_new_input_embeds)
|
178 |
+
new_input_ids.append(cur_new_input_ids)
|
179 |
+
|
180 |
+
input_ids = torch.stack(new_input_ids, dim=0)
|
181 |
+
attention_mask=input_ids.ne(audio_config.llm_pad_token_id)
|
182 |
+
inputs_embeds = torch.stack(new_input_embeds, dim=0)
|
183 |
+
|
184 |
+
batch_label_shift = []
|
185 |
+
batch_speech_pos=[]
|
186 |
+
for i in range(0, len(label_shift), audio_turn):
|
187 |
+
batch_label_shift.append(label_shift[i:i+audio_turn])
|
188 |
+
batch_speech_pos.append(speech_pos[i:i+audio_turn])
|
189 |
+
else:
|
190 |
+
# Inference mode with batch_size=1
|
191 |
+
assert input_ids.shape[0] == 1, "This implementation only supports batch_size=1 during inference"
|
192 |
+
|
193 |
+
# Get all audio token positions in this sample
|
194 |
+
audio_start_token_positions = torch.where(input_ids[0] == audio_config.audio_patch_token)[0]
|
195 |
+
|
196 |
+
# Initialize with original embeddings
|
197 |
+
current_embeds = inputs_embeds[0] # [seq_len, embed_dim]
|
198 |
+
current_ids = input_ids[0] # [seq_len]
|
199 |
+
|
200 |
+
# Process each audio token position sequentially
|
201 |
+
position_shift = 0 # Track position changes due to expansions
|
202 |
+
|
203 |
+
# Ensure shrink_features is properly formatted
|
204 |
+
if isinstance(shrink_features[0], list):
|
205 |
+
# If it's a list of lists (batch_size=1 but multiple turns), flatten it
|
206 |
+
shrink_features = [item for sublist in shrink_features for item in sublist]
|
207 |
+
|
208 |
+
for pos_idx, audio_pos in enumerate(audio_start_token_positions):
|
209 |
+
adjusted_pos = audio_pos + position_shift
|
210 |
+
|
211 |
+
# Get corresponding shrink feature (ensure it's a tensor)
|
212 |
+
shrink_feature = shrink_features[pos_idx]
|
213 |
+
if isinstance(shrink_feature, list):
|
214 |
+
shrink_feature = torch.stack(shrink_feature, dim=0)
|
215 |
+
|
216 |
+
v = shrink_feature.shape[0] # Now this should work
|
217 |
+
# print('len: ', v)
|
218 |
+
|
219 |
+
# Expand the input ids and embeddings
|
220 |
+
current_ids = torch.cat([
|
221 |
+
current_ids[:adjusted_pos],
|
222 |
+
current_ids[adjusted_pos:adjusted_pos+1].repeat(v),
|
223 |
+
current_ids[adjusted_pos + num_patches:]
|
224 |
+
], dim=0)
|
225 |
+
|
226 |
+
current_embeds = torch.cat([
|
227 |
+
current_embeds[:adjusted_pos],
|
228 |
+
shrink_feature,
|
229 |
+
current_embeds[adjusted_pos + num_patches:]
|
230 |
+
], dim=0)
|
231 |
+
|
232 |
+
# Update position shift for next iteration
|
233 |
+
position_shift += (v - num_patches)
|
234 |
+
|
235 |
+
# Update the tensors (unsqueeze to restore batch dim)
|
236 |
+
input_ids = current_ids.unsqueeze(0) # [1, new_seq_len]
|
237 |
+
inputs_embeds = current_embeds.unsqueeze(0) # [1, new_seq_len, embed_dim]
|
238 |
+
attention_mask = input_ids.ne(audio_config.llm_pad_token_id)
|
239 |
+
|
240 |
+
# Update inference state tracking
|
241 |
+
if not hasattr(self, 'mask_tensor'):
|
242 |
+
# Initialize with current attention mask
|
243 |
+
self.mask_tensor = attention_mask.clone()
|
244 |
+
self.length = attention_mask.shape[1]
|
245 |
+
else:
|
246 |
+
# Ensure mask tensor is on correct device
|
247 |
+
self.mask_tensor = self.mask_tensor.to(attention_mask.device)
|
248 |
+
|
249 |
+
# Expand mask tensor if needed
|
250 |
+
if self.mask_tensor.shape[1] < attention_mask.shape[1]:
|
251 |
+
new_mask = torch.zeros(1, attention_mask.shape[1],
|
252 |
+
dtype=torch.bool,
|
253 |
+
device=attention_mask.device)
|
254 |
+
new_mask[0, :self.mask_tensor.shape[1]] = self.mask_tensor
|
255 |
+
self.mask_tensor = new_mask
|
256 |
+
|
257 |
+
# Update mask tensor
|
258 |
+
self.mask_tensor[0, :attention_mask.shape[1]] = attention_mask[0]
|
259 |
+
self.length = attention_mask.shape[1]
|
260 |
+
|
261 |
+
attention_mask=self.mask_tensor[:,:self.length]
|
262 |
+
self.length+=1
|
263 |
+
|
264 |
+
return_state=super(ACLlamaModel, self).forward(
|
265 |
+
input_ids=None, attention_mask=attention_mask, past_key_values=past_key_values,
|
266 |
+
inputs_embeds=inputs_embeds, use_cache=use_cache,
|
267 |
+
output_attentions=output_attentions, output_hidden_states=output_hidden_states,
|
268 |
+
return_dict=return_dict
|
269 |
+
)
|
270 |
+
if self.training and audios is not None:
|
271 |
+
return_state["audio_features"] = predict_logits
|
272 |
+
return_state["label_shift"] = batch_label_shift
|
273 |
+
return_state["label_extend"] = label_extend
|
274 |
+
return_state["speech_pos"] = batch_speech_pos
|
275 |
+
#return_state = {"audio_features":predict_logits}
|
276 |
+
return return_state
|
277 |
+
|
278 |
+
|
279 |
+
class ACLlamaForCausalLM(LlamaForCausalLM):
|
280 |
+
config_class = ACLlamaConfig
|
281 |
+
|
282 |
+
def __init__(self, config):
|
283 |
+
super(LlamaForCausalLM, self).__init__(config)
|
284 |
+
self.model = ACLlamaModel(config)
|
285 |
+
|
286 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
287 |
+
|
288 |
+
# t2u by kkq
|
289 |
+
if hasattr(config, "unit_output"):
|
290 |
+
self.unit_output = config.unit_output
|
291 |
+
self.unit_translator = T2ULlamaForCausalLM(config, self.lm_head.weight)
|
292 |
+
|
293 |
+
# Initialize weights and apply final processing
|
294 |
+
self.post_init()
|
295 |
+
|
296 |
+
def get_model(self):
|
297 |
+
return self.model
|
298 |
+
|
299 |
+
def get_unit_translator(self):
|
300 |
+
return self.unit_translator
|
301 |
+
|
302 |
+
def forward(
|
303 |
+
self,
|
304 |
+
input_ids: torch.LongTensor = None,
|
305 |
+
attention_mask: Optional[torch.Tensor] = None,
|
306 |
+
position_ids: Optional[torch.LongTensor] = None,
|
307 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
308 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
309 |
+
labels: Optional[torch.LongTensor] = None,
|
310 |
+
t2u_input_ids: Optional[torch.LongTensor] = None,
|
311 |
+
t2u_labels: Optional[torch.LongTensor] = None,
|
312 |
+
t2u_attention_mask: Optional[torch.Tensor] = None,
|
313 |
+
unit_targets: Optional[torch.Tensor] = None,
|
314 |
+
sub_lengths: Optional[torch.Tensor] = None,
|
315 |
+
asr_targets: Optional[torch.LongTensor] = None,
|
316 |
+
use_cache: Optional[bool] = None,
|
317 |
+
output_attentions: Optional[bool] = None,
|
318 |
+
output_hidden_states: Optional[bool] = None,
|
319 |
+
audios: Optional[torch.FloatTensor] = None,
|
320 |
+
return_dict: Optional[bool] = None,
|
321 |
+
cache_position: Optional[torch.LongTensor] = None,
|
322 |
+
do_task: str = None,
|
323 |
+
assistant_after_audio_shifts: Optional[torch.Tensor] = None,
|
324 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
325 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
326 |
+
output_hidden_states = (
|
327 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
328 |
+
)
|
329 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
330 |
+
|
331 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
332 |
+
|
333 |
+
# t2u by kkq
|
334 |
+
# pretrain(t2u only) finetune(s2t&e2u)
|
335 |
+
do_task = do_task if do_task != None else getattr(self, 'unit_output', None)
|
336 |
+
|
337 |
+
outputs = None
|
338 |
+
hidden_states = None
|
339 |
+
new_shift_labels = None
|
340 |
+
if do_task != "pretrain":
|
341 |
+
outputs = self.model(
|
342 |
+
input_ids=input_ids,
|
343 |
+
attention_mask=attention_mask,
|
344 |
+
past_key_values=past_key_values,
|
345 |
+
inputs_embeds=inputs_embeds,
|
346 |
+
use_cache=use_cache,
|
347 |
+
output_attentions=output_attentions,
|
348 |
+
output_hidden_states=output_hidden_states,
|
349 |
+
return_dict=return_dict,
|
350 |
+
audios=audios
|
351 |
+
)
|
352 |
+
|
353 |
+
|
354 |
+
hidden_states = outputs[0]
|
355 |
+
logits = self.lm_head(hidden_states)
|
356 |
+
|
357 |
+
loss = None
|
358 |
+
if labels is not None and do_task != "pretrain" and do_task != "finetune_kd":
|
359 |
+
if asr_targets is not None:
|
360 |
+
asr_logits = outputs["audio_features"]
|
361 |
+
asr_targets = asr_targets.view(asr_targets.shape[0] * asr_targets.shape[1], asr_targets.shape[2])
|
362 |
+
mask_asr_targets = (asr_targets != IGNORE_TOKEN_ID)
|
363 |
+
target_lengths = mask_asr_targets.sum(1)
|
364 |
+
input_lengths = torch.full(size=(asr_logits.shape[0],), fill_value=asr_logits.shape[1], dtype=torch.long)
|
365 |
+
|
366 |
+
loss_ctc = CTCLoss()
|
367 |
+
|
368 |
+
log_probs = F.log_softmax(asr_logits, dim=-1).transpose(0, 1)
|
369 |
+
#print(asr_targets.shape)
|
370 |
+
#print(input_lengths, target_lengths)
|
371 |
+
|
372 |
+
with torch.backends.cudnn.flags(enabled=False):
|
373 |
+
loss_asr = F.ctc_loss(
|
374 |
+
log_probs,
|
375 |
+
asr_targets,
|
376 |
+
input_lengths,
|
377 |
+
target_lengths,
|
378 |
+
blank=self.model.audio_tower[0].config.audio_patch_token,
|
379 |
+
reduction='mean',
|
380 |
+
zero_infinity=True,
|
381 |
+
)
|
382 |
+
else:
|
383 |
+
loss_asr=0
|
384 |
+
|
385 |
+
shift_labels = labels
|
386 |
+
if "label_shift" in outputs.keys() and len(outputs["label_shift"]) >0:
|
387 |
+
if outputs["label_extend"] != -1:
|
388 |
+
new_shift_labels = torch.full(size=(shift_labels.shape[0], outputs["label_extend"]+shift_labels.shape[1]), fill_value=IGNORE_TOKEN_ID, dtype=torch.long).to(shift_labels.device)
|
389 |
+
for batch in range(len(outputs["label_shift"])):
|
390 |
+
it_lable_shift = outputs["label_shift"][batch]
|
391 |
+
it_speech_pos = outputs["speech_pos"][batch]
|
392 |
+
prefix = 0
|
393 |
+
for i in range(len(it_lable_shift)):
|
394 |
+
if i == len(it_lable_shift) - 1:
|
395 |
+
length = shift_labels.shape[1] - it_speech_pos[i] #len(shift_labels[batch]) - it_speech_pos[i]
|
396 |
+
else:
|
397 |
+
length = it_speech_pos[i + 1] - it_speech_pos[i]
|
398 |
+
prefix += it_lable_shift[i]
|
399 |
+
new_shift_labels[batch][it_speech_pos[i] + prefix: it_speech_pos[i] + length + prefix]= shift_labels[batch][it_speech_pos[i]:it_speech_pos[i]+length]
|
400 |
+
shift_labels = new_shift_labels
|
401 |
+
else:
|
402 |
+
raise NotImplementedError
|
403 |
+
|
404 |
+
# Shift so that tokens < n predict n
|
405 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
406 |
+
shift_labels = shift_labels[..., 1:].contiguous()
|
407 |
+
#print(shift_labels[:,:50])
|
408 |
+
|
409 |
+
#print(shift_labels[:,:150])
|
410 |
+
loss_fct = CrossEntropyLoss()
|
411 |
+
# Flatten the tokens
|
412 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
413 |
+
shift_labels = shift_labels.view(-1)
|
414 |
+
|
415 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
416 |
+
loss = loss_fct(shift_logits, shift_labels)
|
417 |
+
loss = loss + 0.3 * loss_asr
|
418 |
+
|
419 |
+
t2u_output = None
|
420 |
+
if do_task != None and do_task != "skip":
|
421 |
+
if do_task == "finetune_kd":
|
422 |
+
text_start_index = []
|
423 |
+
for batch in range(len(outputs["label_shift"])):
|
424 |
+
text_start_index.append(outputs["speech_pos"][batch][0] + outputs["label_shift"][batch][0]+assistant_after_audio_shifts[batch])
|
425 |
+
|
426 |
+
t2u_embeds_output = self.unit_translator.insert_text_embedding(
|
427 |
+
input_ids=t2u_input_ids,
|
428 |
+
attention_mask=t2u_attention_mask,
|
429 |
+
inputs_embeds=None,
|
430 |
+
labels=t2u_labels,
|
431 |
+
text_labels=labels,
|
432 |
+
shift_text_labels=new_shift_labels,
|
433 |
+
shift_text_hidden_states=hidden_states,
|
434 |
+
unit_targets=unit_targets,
|
435 |
+
sub_lengths=sub_lengths,
|
436 |
+
text_start_index=text_start_index,
|
437 |
+
do_task=do_task,
|
438 |
+
)
|
439 |
+
|
440 |
+
vae_loss, t2u_inputs_embeds, unit_targets, t2u_attention_mask = t2u_embeds_output
|
441 |
+
|
442 |
+
t2u_output = self.unit_translator(
|
443 |
+
input_ids=None,
|
444 |
+
attention_mask=t2u_attention_mask,
|
445 |
+
past_key_values=past_key_values,
|
446 |
+
inputs_embeds=t2u_inputs_embeds,
|
447 |
+
use_cache=use_cache,
|
448 |
+
labels=unit_targets,
|
449 |
+
output_attentions=output_attentions,
|
450 |
+
output_hidden_states=output_hidden_states,
|
451 |
+
return_dict=return_dict,
|
452 |
+
)
|
453 |
+
else:
|
454 |
+
t2u_embeds_output = self.unit_translator.insert_text_embedding(
|
455 |
+
input_ids=t2u_input_ids,
|
456 |
+
attention_mask=t2u_attention_mask,
|
457 |
+
inputs_embeds=None,
|
458 |
+
labels=t2u_labels,
|
459 |
+
text_labels=labels,
|
460 |
+
shift_text_labels=new_shift_labels,
|
461 |
+
shift_text_hidden_states=hidden_states,
|
462 |
+
do_task=do_task,
|
463 |
+
)
|
464 |
+
vae_loss, t2u_inputs_embeds = t2u_embeds_output
|
465 |
+
|
466 |
+
t2u_output = self.unit_translator(
|
467 |
+
input_ids=None,
|
468 |
+
attention_mask=t2u_attention_mask,
|
469 |
+
past_key_values=past_key_values,
|
470 |
+
inputs_embeds=t2u_inputs_embeds,
|
471 |
+
use_cache=use_cache,
|
472 |
+
labels=t2u_labels,
|
473 |
+
output_attentions=output_attentions,
|
474 |
+
output_hidden_states=output_hidden_states,
|
475 |
+
return_dict=return_dict,
|
476 |
+
)
|
477 |
+
t2u_loss = t2u_output[0]
|
478 |
+
# print(do_task, t2u_loss, vae_loss)
|
479 |
+
if vae_loss != None:
|
480 |
+
target_scale = t2u_loss.item() * 0.2
|
481 |
+
vae_loss_weight = target_scale / vae_loss.item() if vae_loss > target_scale else 1.0
|
482 |
+
t2u_loss = t2u_loss + vae_loss_weight * vae_loss
|
483 |
+
#print(vae_loss)
|
484 |
+
|
485 |
+
if loss != None: # S2T + T2U loss
|
486 |
+
# ignore LLM loss
|
487 |
+
# t2u_output["loss"] = t2u_loss
|
488 |
+
# return t2u_output
|
489 |
+
# original version
|
490 |
+
assert do_task in ["finetune"]
|
491 |
+
if loss.item() < 1.0: # 1.7
|
492 |
+
loss = 0.2 * loss + t2u_loss * 2.0
|
493 |
+
else:
|
494 |
+
loss = loss + t2u_loss
|
495 |
+
else:
|
496 |
+
assert do_task in ["pretrain", "finetune_kd"]
|
497 |
+
t2u_output["loss"] = t2u_loss
|
498 |
+
return t2u_output
|
499 |
+
|
500 |
+
#return CausalLMOutputWithPast(
|
501 |
+
# loss=loss,
|
502 |
+
# logits=outputs["audio_features"],
|
503 |
+
#)
|
504 |
+
|
505 |
+
if not return_dict:
|
506 |
+
output = (logits,) + outputs[1:]
|
507 |
+
return (loss,) + output if loss is not None else output
|
508 |
+
|
509 |
+
return CausalLMOutputWithPast(
|
510 |
+
loss=loss,
|
511 |
+
logits=logits,
|
512 |
+
past_key_values=outputs.past_key_values,
|
513 |
+
hidden_states=outputs.hidden_states,
|
514 |
+
attentions=outputs.attentions,
|
515 |
+
)
|
516 |
+
|
517 |
+
def prepare_inputs_for_generation(
|
518 |
+
self,
|
519 |
+
input_ids,
|
520 |
+
past_key_values=None,
|
521 |
+
attention_mask=None,
|
522 |
+
inputs_embeds=None,
|
523 |
+
cache_position=None,
|
524 |
+
position_ids=None,
|
525 |
+
use_cache=True,
|
526 |
+
**kwargs,
|
527 |
+
):
|
528 |
+
# If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
|
529 |
+
# Exception 1: when passing input_embeds, input_ids may be missing entries
|
530 |
+
# Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
|
531 |
+
if past_key_values is not None:
|
532 |
+
if inputs_embeds is not None: # Exception 1
|
533 |
+
input_ids = input_ids[:, -cache_position.shape[0] :]
|
534 |
+
elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2)
|
535 |
+
input_ids = input_ids[:, cache_position]
|
536 |
+
|
537 |
+
if attention_mask is not None and position_ids is None:
|
538 |
+
# create position_ids on the fly for batch generation
|
539 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
540 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
541 |
+
if past_key_values:
|
542 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
543 |
+
|
544 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
545 |
+
if inputs_embeds is not None and cache_position[0] == 0:
|
546 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
547 |
+
else:
|
548 |
+
model_inputs = {"input_ids": input_ids.contiguous()} # `contiguous()` needed for compilation use cases
|
549 |
+
|
550 |
+
model_inputs.update(
|
551 |
+
{
|
552 |
+
"position_ids": position_ids,
|
553 |
+
"cache_position": cache_position,
|
554 |
+
"past_key_values": past_key_values,
|
555 |
+
"use_cache": use_cache,
|
556 |
+
"attention_mask": attention_mask,
|
557 |
+
}
|
558 |
+
)
|
559 |
+
model_inputs.update({"audios": kwargs["audios"]} if "audios" in kwargs.keys() else {})
|
560 |
+
model_inputs.update({"do_task": kwargs["do_task"]} if "do_task" in kwargs.keys() else {})
|
561 |
+
model_inputs.update({"return_dict": kwargs["return_dict_in_generate"]} if "return_dict_in_generate" in kwargs.keys() else {})
|
562 |
+
return model_inputs
|
563 |
+
|
564 |
+
AutoConfig.register("ACLlama", ACLlamaConfig)
|
565 |
+
AutoModelForCausalLM.register(ACLlamaConfig, ACLlamaForCausalLM)
|
EchoX-Vocoder/checkpoint_last.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:abb7b49cc59bbf058719cdae2252069dce1bb7b73362a0ec5273670ed7a6d4cc
|
3 |
+
size 389348172
|
EchoX-Vocoder/config.json
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"input_wavs_dir": "/private/home/adampolyak/datasets/LJ/LJSpeech-1.1/wavs_16khz_padded",
|
3 |
+
"input_training_file": "/large_experiments/ust/annl/datasets/tts/LJSpeech/filelist/mhubert_vp_en_es_fr_it3_400k/lj_train_layer11_hubert1000_filelist.txt",
|
4 |
+
"input_validation_file": "/large_experiments/ust/annl/datasets/tts/LJSpeech/filelist/mhubert_vp_en_es_fr_it3_400k/lj_dev_layer11_hubert1000_filelist.txt",
|
5 |
+
|
6 |
+
"resblock": "1",
|
7 |
+
"num_gpus": 0,
|
8 |
+
"batch_size": 16,
|
9 |
+
"learning_rate": 0.0002,
|
10 |
+
"adam_b1": 0.8,
|
11 |
+
"adam_b2": 0.99,
|
12 |
+
"lr_decay": 0.999,
|
13 |
+
"seed": 1234,
|
14 |
+
|
15 |
+
"upsample_rates": [5,4,4,2,2],
|
16 |
+
"upsample_kernel_sizes": [11,8,8,4,4],
|
17 |
+
"upsample_initial_channel": 512,
|
18 |
+
"resblock_kernel_sizes": [3,7,11],
|
19 |
+
"resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
|
20 |
+
"num_embeddings": 1000,
|
21 |
+
"embedding_dim": 128,
|
22 |
+
"model_in_dim": 128,
|
23 |
+
|
24 |
+
"segment_size": 8960,
|
25 |
+
"code_hop_size": 320,
|
26 |
+
"f0": false,
|
27 |
+
"num_mels": 80,
|
28 |
+
"num_freq": 1025,
|
29 |
+
"n_fft": 1024,
|
30 |
+
"hop_size": 256,
|
31 |
+
"win_size": 1024,
|
32 |
+
|
33 |
+
"dur_prediction_weight": 1.0,
|
34 |
+
"dur_predictor_params": {
|
35 |
+
"encoder_embed_dim": 128,
|
36 |
+
"var_pred_hidden_dim": 128,
|
37 |
+
"var_pred_kernel_size": 3,
|
38 |
+
"var_pred_dropout": 0.5
|
39 |
+
},
|
40 |
+
|
41 |
+
"sampling_rate": 16000,
|
42 |
+
|
43 |
+
"fmin": 0,
|
44 |
+
"fmax": 8000,
|
45 |
+
"fmax_for_loss": null,
|
46 |
+
|
47 |
+
"num_workers": 4,
|
48 |
+
|
49 |
+
"dist_config": {
|
50 |
+
"dist_backend": "nccl",
|
51 |
+
"dist_url": "env://"
|
52 |
+
}
|
53 |
+
}
|
EchoX-Vocoder/g_00500000
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0d1f7188b95b06304bc05e524fddf93c7fe682fdd93acff022685663a5e26b97
|
3 |
+
size 54051213
|
EchoX-Vocoder/re_config.log
ADDED
File without changes
|
EchoX-Vocoder/spm_1k.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d95d5585291329feaf35d3cb39fe5181e4987549097a9daa36f468dab9e82556
|
3 |
+
size 254653
|
Echox_copy_stream.py
ADDED
@@ -0,0 +1,426 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys
|
2 |
+
from ACLlama_el_s2s import ACLlamaForCausalLM
|
3 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoConfig, WhisperProcessor
|
4 |
+
from peft import PeftModel, PeftConfig
|
5 |
+
import json
|
6 |
+
from tqdm import tqdm
|
7 |
+
import torch
|
8 |
+
import re
|
9 |
+
import os
|
10 |
+
torch.backends.cudnn.benchmark = False
|
11 |
+
import librosa
|
12 |
+
from text_to_speech import *
|
13 |
+
import torch.nn.functional as F
|
14 |
+
from concurrent.futures import ThreadPoolExecutor, as_completed
|
15 |
+
|
16 |
+
from transformers import logging as hf_logging
|
17 |
+
hf_logging.set_verbosity_error()
|
18 |
+
from huggingface_hub import hf_hub_download
|
19 |
+
from typing import Dict, Optional, List
|
20 |
+
import tempfile
|
21 |
+
import select
|
22 |
+
from copy import deepcopy
|
23 |
+
from typing import Generator, Tuple
|
24 |
+
|
25 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "true"
|
26 |
+
|
27 |
+
def load_model(args, device):
|
28 |
+
quantization_config = None
|
29 |
+
hf_token = os.getenv("HF_TOKEN")
|
30 |
+
|
31 |
+
# load based model
|
32 |
+
model = ACLlamaForCausalLM.from_pretrained(
|
33 |
+
args.base_model_path,
|
34 |
+
device_map=None,
|
35 |
+
torch_dtype=torch.float16,
|
36 |
+
quantization_config=quantization_config,
|
37 |
+
token=hf_token,
|
38 |
+
).eval().to(device)
|
39 |
+
for module in model.model.audio_tower:
|
40 |
+
module = module.to(device)
|
41 |
+
|
42 |
+
if args.peft_model_id:
|
43 |
+
lora_config = PeftConfig.from_pretrained(args.peft_model_id)
|
44 |
+
torch.cuda.empty_cache()
|
45 |
+
model = PeftModel.from_pretrained(model, args.peft_model_id, config=lora_config).to(
|
46 |
+
dtype=torch.float16, device=device
|
47 |
+
)
|
48 |
+
model = model.merge_and_unload()
|
49 |
+
|
50 |
+
model.eval()
|
51 |
+
|
52 |
+
# load tokenizer
|
53 |
+
tokenizer = AutoTokenizer.from_pretrained(args.base_model_path, token=hf_token)
|
54 |
+
|
55 |
+
audio_config = model.get_model().audio_tower[0].config
|
56 |
+
audio_config.audio_patch_token = tokenizer.get_vocab()["<audio_patch>"]
|
57 |
+
audio_config.llm_pad_token_id = tokenizer.pad_token_id
|
58 |
+
audio_config.audio_patch_size = args.audio_token_len
|
59 |
+
|
60 |
+
|
61 |
+
# whisper processor
|
62 |
+
audio_processor = WhisperProcessor.from_pretrained(args.audio_tower, torch_dtype=torch.float16)
|
63 |
+
|
64 |
+
# t2u
|
65 |
+
unit_translator = model.get_unit_translator().eval()
|
66 |
+
return model, audio_processor, tokenizer, unit_translator
|
67 |
+
|
68 |
+
def load_speech_model(device):
|
69 |
+
vocoder = "./EchoX-Vocoder/g_00500000"
|
70 |
+
vocoder_cfg = "./EchoX-Vocoder/config.json"
|
71 |
+
voc_cfg = get_vocoder_config(vocoder, vocoder_cfg)
|
72 |
+
vocoder = load_units_vocoder(voc_cfg, device)
|
73 |
+
return vocoder, voc_cfg
|
74 |
+
|
75 |
+
# def load_speech_model(device):
|
76 |
+
# hf_token = os.getenv("HF_TOKEN")
|
77 |
+
|
78 |
+
# vocoder_repo_id = "FreedomIntelligence/EchoX-Vocoder"
|
79 |
+
|
80 |
+
# cache_path = './hf_cache'
|
81 |
+
# vocoder_path = hf_hub_download(repo_id=vocoder_repo_id, filename="g_00500000", token=hf_token, cache_dir=cache_path)
|
82 |
+
# vocoder_cfg_path = hf_hub_download(repo_id=vocoder_repo_id, filename="config.json", token=hf_token, cache_dir=cache_path)
|
83 |
+
|
84 |
+
# voc_cfg = get_vocoder_config(vocoder_path, vocoder_cfg_path)
|
85 |
+
# vocoder = load_units_vocoder(voc_cfg, device)
|
86 |
+
# return vocoder, voc_cfg
|
87 |
+
|
88 |
+
class EchoxAssistant():
|
89 |
+
def __init__(self):
|
90 |
+
class BasicSetting:
|
91 |
+
def __init__(self):
|
92 |
+
self.device = "cuda:0"
|
93 |
+
self.sampling_rate = 16000
|
94 |
+
self.audio_token_len = 1 # 1500 = 300 token x 5 compress
|
95 |
+
self.stop = "</s>"
|
96 |
+
self.base_model_path = "FreedomIntelligence/EchoX-8B"
|
97 |
+
self.peft_model_id = None
|
98 |
+
self.audio_tower = "openai/whisper-large-v3"
|
99 |
+
self.args = BasicSetting()
|
100 |
+
self.device = "cuda"
|
101 |
+
self.vocoder, self.voc_cfg= load_speech_model(self.device)
|
102 |
+
self.model, self.audio_processor, self.tokenizer, self.unit_translator = load_model(self.args, self.device)
|
103 |
+
self.audio_executor = ThreadPoolExecutor(max_workers=2)
|
104 |
+
# self.specAug = SpecAugmentTransform()
|
105 |
+
# special_token
|
106 |
+
DEFAULT_AUDIO_PATCH_TOKEN = "<audio_patch>"
|
107 |
+
audio_placeholder = DEFAULT_AUDIO_PATCH_TOKEN * self.args.audio_token_len
|
108 |
+
audio_placeholder = "\n"+audio_placeholder
|
109 |
+
self.audio_placeholder_ids = self.tokenizer(audio_placeholder).input_ids
|
110 |
+
|
111 |
+
self.begin_of_text_id = self.tokenizer.get_vocab()["<|begin_of_text|>"]
|
112 |
+
self.start_header_id = self.tokenizer.get_vocab()["<|start_header_id|>"]
|
113 |
+
self.end_header_id = self.tokenizer.get_vocab()["<|end_header_id|>"]
|
114 |
+
self.eot_id = self.tokenizer.get_vocab()["<|eot_id|>"]
|
115 |
+
self.nl_tokens = self.tokenizer('\n').input_ids
|
116 |
+
self._system = self.tokenizer('system').input_ids
|
117 |
+
self._user = self.tokenizer('user').input_ids
|
118 |
+
self._assistant = self.tokenizer('assistant').input_ids
|
119 |
+
self._speaker = self.tokenizer('speaker').input_ids
|
120 |
+
|
121 |
+
self.max_len = 1024
|
122 |
+
self.unit_max_len = 2048
|
123 |
+
self.system_message = "You are a helpful language and speech assistant. You are able to understand the speech content that the user provides, and assist the user with a variety of tasks using natural language."
|
124 |
+
|
125 |
+
def _generate_audio_segment(self, segment_hidden_states):
|
126 |
+
try:
|
127 |
+
audio_units = self._generate_audio_units_from_hidden_states(segment_hidden_states)
|
128 |
+
if audio_units:
|
129 |
+
audio_float32 = self.generate_with_speech_model([list(map(int, audio_units.split(" ")))])
|
130 |
+
audio_int16 = (audio_float32 * 32767).astype(np.int16)
|
131 |
+
|
132 |
+
print(f"Generated audio segment in background: {len(audio_units.split())} units")
|
133 |
+
return (16000, audio_int16)
|
134 |
+
return None
|
135 |
+
except Exception as e:
|
136 |
+
print(f"Background audio generation error: {e}")
|
137 |
+
return None
|
138 |
+
|
139 |
+
def gen_model_inputs(
|
140 |
+
self,
|
141 |
+
sources,
|
142 |
+
tokenizer,
|
143 |
+
max_len,
|
144 |
+
system_message,
|
145 |
+
audio_placeholder_ids, begin_of_text_id, start_header_id, end_header_id, eot_id, nl_tokens, _system, _user, _assistant,
|
146 |
+
) -> dict:
|
147 |
+
# max_len 512
|
148 |
+
|
149 |
+
# Apply prompt templates
|
150 |
+
input_ids, audio_paths = [], []
|
151 |
+
audio_path = []
|
152 |
+
|
153 |
+
for source in sources:
|
154 |
+
input_id = []
|
155 |
+
system = [begin_of_text_id] + [start_header_id] + _system + [end_header_id] + nl_tokens + tokenizer(system_message).input_ids + [eot_id]
|
156 |
+
input_id += system
|
157 |
+
|
158 |
+
for j, item in enumerate(source["conversations"]):
|
159 |
+
role = item["from"]
|
160 |
+
value = item["value"]
|
161 |
+
_audio_path = None
|
162 |
+
|
163 |
+
if role == 'user':
|
164 |
+
if "audio" in item.keys():
|
165 |
+
_input_id = [start_header_id] + _user + [end_header_id] + audio_placeholder_ids + tokenizer(value).input_ids + [eot_id]
|
166 |
+
_audio_path = item["audio"]
|
167 |
+
else:
|
168 |
+
_input_id = [start_header_id] + _user + [end_header_id] + tokenizer(value).input_ids + [eot_id]
|
169 |
+
|
170 |
+
elif role == 'assistant':
|
171 |
+
_input_id = [start_header_id] + _assistant + [end_header_id] + nl_tokens + tokenizer(value).input_ids + [eot_id]
|
172 |
+
|
173 |
+
else:
|
174 |
+
raise NotImplementedError
|
175 |
+
input_id += _input_id
|
176 |
+
|
177 |
+
if _audio_path:
|
178 |
+
audio_path.append(_audio_path)
|
179 |
+
assistant_input_id = [start_header_id] + _assistant + [end_header_id] + nl_tokens
|
180 |
+
input_id += assistant_input_id
|
181 |
+
|
182 |
+
audio_num = int(input_id.count(audio_placeholder_ids[-1]) / self.args.audio_token_len)
|
183 |
+
assert len(audio_path) == audio_num
|
184 |
+
if len(input_id) >= max_len:
|
185 |
+
print(f"[WARNING] Your Input Length More Than {max_len}")
|
186 |
+
input_ids.append(input_id[:max_len])
|
187 |
+
audio_paths.append(audio_path)
|
188 |
+
input_ids = torch.tensor(input_ids, dtype=torch.int)
|
189 |
+
return dict(
|
190 |
+
input_ids=input_ids,
|
191 |
+
audio_paths=audio_paths,
|
192 |
+
attention_mask=input_ids.ne(tokenizer.pad_token_id),
|
193 |
+
)
|
194 |
+
|
195 |
+
def get_unit_result(self, ret):
|
196 |
+
# print(ret)
|
197 |
+
self.unit_translator.generation_config.pad_token_id = self.tokenizer.eos_token_id
|
198 |
+
input_ids = ret["input_ids"]
|
199 |
+
ret["input_ids"] = None
|
200 |
+
model_outputs = self.unit_translator.generate(
|
201 |
+
**ret,
|
202 |
+
max_new_tokens=2048,
|
203 |
+
eos_token_id=self.tokenizer.eos_token_id,
|
204 |
+
)
|
205 |
+
# print(model_outputs, model_outputs.shape)
|
206 |
+
output_ids = model_outputs
|
207 |
+
unit_output = self.tokenizer.batch_decode(output_ids)[0]
|
208 |
+
if "▁" in unit_output:
|
209 |
+
unit_output = ''.join(re.findall(r"<\|unit_(.*?)\|>", unit_output))
|
210 |
+
|
211 |
+
units = re.findall(r'\d+', unit_output)
|
212 |
+
|
213 |
+
#TODO grid of unk unit
|
214 |
+
new_units = []
|
215 |
+
for unit in units:
|
216 |
+
if int(unit) < 1000:
|
217 |
+
new_units.append(unit)
|
218 |
+
|
219 |
+
units = ' '.join(new_units)
|
220 |
+
return units
|
221 |
+
|
222 |
+
|
223 |
+
def _inference(
|
224 |
+
self,
|
225 |
+
prompt,
|
226 |
+
**kwargs,
|
227 |
+
):
|
228 |
+
audio_paths = []
|
229 |
+
response = []
|
230 |
+
for item in prompt:
|
231 |
+
for conv in item["conversations"]:
|
232 |
+
if "audio" in conv:
|
233 |
+
audio_paths.append(conv["audio"])
|
234 |
+
|
235 |
+
model_inputs = self.gen_model_inputs(
|
236 |
+
prompt,
|
237 |
+
self.tokenizer,
|
238 |
+
self.max_len,
|
239 |
+
self.system_message,
|
240 |
+
self.audio_placeholder_ids, self.begin_of_text_id, self.start_header_id, self.end_header_id, self.eot_id, self.nl_tokens, self._system, self._user, self._assistant)
|
241 |
+
|
242 |
+
audio_list = []
|
243 |
+
if audio_paths and audio_paths[0] is not None:
|
244 |
+
for audio_path in audio_paths:
|
245 |
+
# print("read audio file name: ", audio_path)
|
246 |
+
audio, _ = librosa.load(audio_path, sr=self.args.sampling_rate)
|
247 |
+
audio_feat = self.audio_processor(audio, sampling_rate=self.args.sampling_rate, return_tensors="pt").input_features
|
248 |
+
audio_list.append(audio_feat)
|
249 |
+
audio_feats = torch.stack(audio_list, dim=0)
|
250 |
+
audio_feats = audio_feats.to(dtype=torch.float16).to(self.device)
|
251 |
+
|
252 |
+
if not audio_list:
|
253 |
+
ret = dict(
|
254 |
+
input_ids=model_inputs["input_ids"].to(self.device),
|
255 |
+
attention_mask=model_inputs["attention_mask"].to(self.device),
|
256 |
+
)
|
257 |
+
else:
|
258 |
+
ret = dict(
|
259 |
+
input_ids=model_inputs["input_ids"].to(self.device),
|
260 |
+
attention_mask=model_inputs["attention_mask"].to(self.device),
|
261 |
+
audios=audio_feats,
|
262 |
+
)
|
263 |
+
|
264 |
+
self.model.generation_config.pad_token_id = self.tokenizer.eos_token_id
|
265 |
+
#print(self.model.lm_head.weight.shape)
|
266 |
+
|
267 |
+
dot_input_ids = self.tokenizer(".", return_tensors="pt").input_ids.to(self.device) # 形状: (1, 2), 值: [[128000, 13]]
|
268 |
+
period_token_id = dot_input_ids[0, -1]
|
269 |
+
period_lm_head_embedding = self.model.lm_head.weight[period_token_id]
|
270 |
+
|
271 |
+
input_ids = ret["input_ids"]
|
272 |
+
attention_mask = ret["attention_mask"]
|
273 |
+
input_token_len = input_ids.shape[1]
|
274 |
+
|
275 |
+
max_new_tokens = kwargs.get('max_new_tokens', 512)
|
276 |
+
temperature = kwargs.get('temperature', 0.2)
|
277 |
+
top_p = kwargs.get('top_p', 0.9)
|
278 |
+
do_sample = kwargs.get('do_sample', True)
|
279 |
+
|
280 |
+
current_text = ""
|
281 |
+
accumulated_hidden_states = []
|
282 |
+
accumulated_tokens = []
|
283 |
+
similarity_scores = []
|
284 |
+
segment_start_idx = 0
|
285 |
+
|
286 |
+
current_input_ids = input_ids
|
287 |
+
current_attention_mask = attention_mask
|
288 |
+
past_key_values = None
|
289 |
+
|
290 |
+
audio_futures = []
|
291 |
+
segmentation_latency = 5
|
292 |
+
|
293 |
+
with torch.no_grad():
|
294 |
+
for step in range(max_new_tokens):
|
295 |
+
while audio_futures and audio_futures[0].done():
|
296 |
+
completed_future = audio_futures.pop(0)
|
297 |
+
audio_data = completed_future.result()
|
298 |
+
if audio_data:
|
299 |
+
yield None, audio_data
|
300 |
+
|
301 |
+
if current_input_ids is None:
|
302 |
+
break
|
303 |
+
|
304 |
+
model_kwargs = {
|
305 |
+
"input_ids": current_input_ids,
|
306 |
+
"attention_mask": current_attention_mask,
|
307 |
+
"past_key_values": past_key_values,
|
308 |
+
"use_cache": True,
|
309 |
+
"output_hidden_states": True,
|
310 |
+
"do_task": "skip"
|
311 |
+
}
|
312 |
+
|
313 |
+
if step == 0 and "audios" in ret:
|
314 |
+
model_kwargs["audios"] = ret["audios"]
|
315 |
+
|
316 |
+
outputs = self.model(**model_kwargs)
|
317 |
+
|
318 |
+
logits = outputs.logits
|
319 |
+
hidden_states = outputs.hidden_states[-1]
|
320 |
+
past_key_values = outputs.past_key_values
|
321 |
+
|
322 |
+
next_token_logits = logits[:, -1, :] # [batch_size, vocab_size]
|
323 |
+
|
324 |
+
if do_sample:
|
325 |
+
next_token_logits = next_token_logits / temperature
|
326 |
+
|
327 |
+
sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
|
328 |
+
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
|
329 |
+
|
330 |
+
sorted_indices_to_remove = cumulative_probs > top_p
|
331 |
+
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
|
332 |
+
sorted_indices_to_remove[..., 0] = 0
|
333 |
+
|
334 |
+
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
|
335 |
+
next_token_logits[indices_to_remove] = float('-inf')
|
336 |
+
|
337 |
+
probs = F.softmax(next_token_logits, dim=-1)
|
338 |
+
next_token = torch.multinomial(probs, num_samples=1)
|
339 |
+
else:
|
340 |
+
next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)
|
341 |
+
|
342 |
+
if next_token.item() == self.tokenizer.eos_token_id:
|
343 |
+
current_input_ids = None
|
344 |
+
continue
|
345 |
+
|
346 |
+
accumulated_tokens.append(next_token.item())
|
347 |
+
last_hidden_state = hidden_states[0, -1] # [hidden_dim]
|
348 |
+
accumulated_hidden_states.append(last_hidden_state)
|
349 |
+
|
350 |
+
similarity = F.cosine_similarity(last_hidden_state, period_lm_head_embedding, dim=0).item()
|
351 |
+
similarity_scores.append(similarity)
|
352 |
+
|
353 |
+
token_text = self.tokenizer.decode([next_token.item()], skip_special_tokens=True)
|
354 |
+
current_text += token_text
|
355 |
+
|
356 |
+
yield current_text, None
|
357 |
+
|
358 |
+
current_idx = len(similarity_scores) - 1
|
359 |
+
check_idx = current_idx - segmentation_latency
|
360 |
+
if check_idx >= 0:
|
361 |
+
similarity_at_check = similarity_scores[check_idx]
|
362 |
+
is_peak = self._is_local_maximum(similarity_scores, check_idx, window=segmentation_latency)
|
363 |
+
should_segment = (is_peak and
|
364 |
+
check_idx - segment_start_idx >= 50) or (
|
365 |
+
is_peak and
|
366 |
+
similarity_at_check > 0.1 and
|
367 |
+
check_idx - segment_start_idx >= 20
|
368 |
+
)
|
369 |
+
|
370 |
+
if should_segment:
|
371 |
+
segment_end_idx = check_idx + 1
|
372 |
+
print(f"Segmenting at step {segment_end_idx-1}, similarity={similarity_at_check:.4f}. Submitting to background audio generation.")
|
373 |
+
|
374 |
+
segment_hidden_states = torch.stack(
|
375 |
+
accumulated_hidden_states[segment_start_idx:segment_end_idx], dim=0
|
376 |
+
).unsqueeze(0)
|
377 |
+
|
378 |
+
future = self.audio_executor.submit(self._generate_audio_segment, segment_hidden_states)
|
379 |
+
audio_futures.append(future)
|
380 |
+
|
381 |
+
segment_start_idx = segment_end_idx
|
382 |
+
|
383 |
+
current_input_ids = next_token
|
384 |
+
current_attention_mask = torch.ones_like(next_token)
|
385 |
+
|
386 |
+
if segment_start_idx < len(accumulated_hidden_states):
|
387 |
+
print(f"Processing final segment from {segment_start_idx} to {len(accumulated_hidden_states)}")
|
388 |
+
segment_hidden_states = torch.stack(
|
389 |
+
accumulated_hidden_states[segment_start_idx:], dim=0
|
390 |
+
).unsqueeze(0)
|
391 |
+
future = self.audio_executor.submit(self._generate_audio_segment, segment_hidden_states)
|
392 |
+
audio_futures.append(future)
|
393 |
+
|
394 |
+
for future in audio_futures:
|
395 |
+
audio_data = future.result()
|
396 |
+
if audio_data:
|
397 |
+
yield None, audio_data
|
398 |
+
|
399 |
+
def _is_local_maximum(self, scores, idx, window=5):
|
400 |
+
start = max(0, idx - window)
|
401 |
+
end = min(len(scores), idx + window + 1)
|
402 |
+
local_scores = scores[start:end]
|
403 |
+
return scores[idx] == max(local_scores)
|
404 |
+
|
405 |
+
def _generate_audio_units_from_hidden_states(self, hidden_states):
|
406 |
+
try:
|
407 |
+
_, adapted_inputs_embeds = self.unit_translator.insert_text_embedding(
|
408 |
+
inputs_embeds=hidden_states,
|
409 |
+
do_task="skip",
|
410 |
+
)
|
411 |
+
|
412 |
+
attention_mask = torch.ones(adapted_inputs_embeds.shape[:2]).to(self.device)
|
413 |
+
ret = dict(
|
414 |
+
input_ids=None,
|
415 |
+
inputs_embeds=adapted_inputs_embeds,
|
416 |
+
attention_mask=attention_mask,
|
417 |
+
)
|
418 |
+
|
419 |
+
return self.get_unit_result(ret)
|
420 |
+
except Exception as e:
|
421 |
+
print(f"Error generating audio units: {e}")
|
422 |
+
return None
|
423 |
+
|
424 |
+
def generate_with_speech_model(self, units):
|
425 |
+
wav = gen_wav(self.vocoder, self.voc_cfg, units, self.device)
|
426 |
+
return wav
|
T2ULlama_CR_online.py
ADDED
@@ -0,0 +1,412 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional, Tuple, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
from torch.nn import CrossEntropyLoss, CTCLoss
|
7 |
+
import transformers
|
8 |
+
|
9 |
+
from transformers import AutoConfig, AutoModelForCausalLM, \
|
10 |
+
LlamaConfig, LlamaModel, LlamaForCausalLM
|
11 |
+
from transformers.trainer_pt_utils import LabelSmoother
|
12 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
|
13 |
+
from transformers import (
|
14 |
+
WhisperProcessor,
|
15 |
+
WhisperModel,
|
16 |
+
)
|
17 |
+
|
18 |
+
IGNORE_TOKEN_ID = LabelSmoother.ignore_index
|
19 |
+
|
20 |
+
|
21 |
+
def padding_tensor(tensor, length, dim=0, pad=False):
|
22 |
+
|
23 |
+
if length == 0:
|
24 |
+
return tensor
|
25 |
+
|
26 |
+
assert length > 0, f"Wrong padding length: {length}"
|
27 |
+
|
28 |
+
shape = list(tensor.shape)
|
29 |
+
assert dim < len(shape), f"dim {dim} out of shape {shape}"
|
30 |
+
shape[dim] = length
|
31 |
+
padding_tensor = torch.cat(
|
32 |
+
(
|
33 |
+
tensor,
|
34 |
+
torch.full(tuple(shape), pad, dtype=tensor.dtype, device=tensor.device)
|
35 |
+
),
|
36 |
+
dim=dim
|
37 |
+
)
|
38 |
+
return padding_tensor
|
39 |
+
|
40 |
+
|
41 |
+
class T2ULlamaConfig(LlamaConfig):
|
42 |
+
model_type = "T2ULlama"
|
43 |
+
|
44 |
+
class T2ULlamaForCausalLM(LlamaForCausalLM):
|
45 |
+
config_class = T2ULlamaConfig
|
46 |
+
|
47 |
+
def __init__(self, config, embedding_weight=None):
|
48 |
+
|
49 |
+
self.current_step = 0
|
50 |
+
self.log = {}
|
51 |
+
|
52 |
+
super(LlamaForCausalLM, self).__init__(config)
|
53 |
+
self.config = config
|
54 |
+
self.training_stage = config.unit_output
|
55 |
+
self.pad_token_id = 128009
|
56 |
+
|
57 |
+
llama_config = T2ULlamaConfig(**config.to_dict(),
|
58 |
+
batch_first=True,
|
59 |
+
norm_first=True
|
60 |
+
)
|
61 |
+
llama_config.architectures = ["T2ULlamaForCausalLM"]
|
62 |
+
llama_config.pad_token_id = self.pad_token_id
|
63 |
+
llama_config.vocab_size += llama_config.unit_vocab_size
|
64 |
+
#######################################################
|
65 |
+
llama_config.unit_model = "medium"
|
66 |
+
llama_config.max_position_embeddings = 2048 # 1024 1536 2048 # origin 1024 reduced 512
|
67 |
+
#######################################################
|
68 |
+
if hasattr(llama_config, "unit_model"):
|
69 |
+
if llama_config.unit_model == "large":
|
70 |
+
llama_config.num_hidden_layers = 2
|
71 |
+
# llama_config.hidden_size = 4096
|
72 |
+
# llama_config.num_attention_heads = 32
|
73 |
+
# llama_config.intermediate_size = 14336
|
74 |
+
# llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads
|
75 |
+
|
76 |
+
elif llama_config.unit_model == "tiny":
|
77 |
+
llama_config.num_hidden_layers = 4
|
78 |
+
llama_config.hidden_size = 512
|
79 |
+
llama_config.num_attention_heads = 8
|
80 |
+
llama_config.intermediate_size = 2048
|
81 |
+
llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads
|
82 |
+
else:
|
83 |
+
llama_config.num_hidden_layers = 8
|
84 |
+
llama_config.hidden_size = 768
|
85 |
+
llama_config.num_attention_heads = 12
|
86 |
+
llama_config.num_key_value_heads = 12
|
87 |
+
llama_config.intermediate_size = 2048
|
88 |
+
llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads
|
89 |
+
else:
|
90 |
+
llama_config.num_hidden_layers = 6
|
91 |
+
llama_config.hidden_size = 512
|
92 |
+
llama_config.num_attention_heads = 8
|
93 |
+
llama_config.intermediate_size = 2048
|
94 |
+
llama_config.head_dim = llama_config.hidden_size // llama_config.num_attention_heads
|
95 |
+
# print(llama_config)
|
96 |
+
|
97 |
+
self.model = LlamaModel(llama_config)
|
98 |
+
# share embedding 0501 by kkq
|
99 |
+
self.model.embed_tokens = nn.Embedding(num_embeddings=config.vocab_size, embedding_dim=config.hidden_size, padding_idx=self.pad_token_id) # redefine
|
100 |
+
self.unit_embedding = nn.Linear(config.hidden_size, llama_config.unit_vocab_size, bias=False)
|
101 |
+
self.adapter = nn.Linear(config.hidden_size, llama_config.hidden_size, bias = True)
|
102 |
+
self.lm_head = nn.Linear(llama_config.hidden_size, llama_config.vocab_size, bias=False)
|
103 |
+
|
104 |
+
if self.training_stage == "pretrain":
|
105 |
+
pass
|
106 |
+
elif self.training_stage == "finetune" or self.training_stage == "finetune_kd" or self.training_stage == "finetune_kd_online":
|
107 |
+
self.aligner_MLP = nn.Sequential(
|
108 |
+
nn.Linear(config.hidden_size, config.intermediate_size),
|
109 |
+
nn.ReLU(),
|
110 |
+
nn.Dropout(0.1),
|
111 |
+
nn.Linear(config.intermediate_size, config.hidden_size),
|
112 |
+
)
|
113 |
+
torch.nn.init.ones_(self.aligner_MLP[0].weight)
|
114 |
+
torch.nn.init.zeros_(self.aligner_MLP[0].bias)
|
115 |
+
torch.nn.init.ones_(self.aligner_MLP[3].weight)
|
116 |
+
torch.nn.init.zeros_(self.aligner_MLP[3].bias)
|
117 |
+
|
118 |
+
# Initialize weights and apply final processing
|
119 |
+
self.post_init()
|
120 |
+
|
121 |
+
def get_model(self):
|
122 |
+
return self.model
|
123 |
+
|
124 |
+
def insert_text_embedding(
|
125 |
+
self,
|
126 |
+
input_ids: torch.LongTensor = None,
|
127 |
+
attention_mask: Optional[torch.Tensor] = None,
|
128 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
129 |
+
labels: Optional[torch.LongTensor] = None,
|
130 |
+
text_labels: Optional[torch.LongTensor] = None,
|
131 |
+
shift_text_labels: Optional[torch.LongTensor] = None,
|
132 |
+
shift_text_hidden_states: Optional[torch.FloatTensor] = None,
|
133 |
+
unit_targets: Optional[torch.LongTensor] = None,
|
134 |
+
sub_lengths: Optional[torch.LongTensor] = None,
|
135 |
+
text_start_index: Optional[torch.LongTensor] = None,
|
136 |
+
do_task: str = None,
|
137 |
+
**kwargs: dict,
|
138 |
+
):
|
139 |
+
|
140 |
+
if inputs_embeds == None:
|
141 |
+
# share embedding 0501 by kkq
|
142 |
+
embed_tokens_weight = torch.cat(
|
143 |
+
[
|
144 |
+
self.model.embed_tokens.weight.detach(), self.unit_embedding.weight
|
145 |
+
],
|
146 |
+
dim = 0,
|
147 |
+
)
|
148 |
+
# print(embed_tokens_weight, embed_tokens_weight.shape)
|
149 |
+
inputs_embeds = F.embedding(input_ids, embed_tokens_weight, padding_idx=self.pad_token_id)
|
150 |
+
|
151 |
+
emb_loss = None
|
152 |
+
if do_task == "pretrain":
|
153 |
+
if self.training:
|
154 |
+
if hasattr(self, "embedding_dropout"):
|
155 |
+
emb_origin_mask = text_labels != -100
|
156 |
+
origin_padding_length = labels.shape[-1] - emb_origin_mask.shape[-1]
|
157 |
+
extend_emb_origin_mask = padding_tensor(emb_origin_mask, origin_padding_length, 1, False)
|
158 |
+
extend_emb_origin_mask = ~extend_emb_origin_mask.unsqueeze(-1).expand_as(inputs_embeds)
|
159 |
+
|
160 |
+
# Π-Model + noise
|
161 |
+
log_var = self.perturb(inputs_embeds)
|
162 |
+
perturbed_inputs_embeds_2 = inputs_embeds + torch.randn_like(inputs_embeds) * (torch.exp(0.5 * log_var) + 1e-6)
|
163 |
+
# Π-Model + dropout
|
164 |
+
perturbed_inputs_embeds_1 = self.embedding_dropout(inputs_embeds)
|
165 |
+
perturbed_inputs_embeds_2 = self.embedding_dropout(perturbed_inputs_embeds_2)
|
166 |
+
perturbed_inputs_embeds_1 = torch.where(extend_emb_origin_mask, inputs_embeds, perturbed_inputs_embeds_1)
|
167 |
+
perturbed_inputs_embeds_2 = torch.where(extend_emb_origin_mask, inputs_embeds, perturbed_inputs_embeds_2)
|
168 |
+
|
169 |
+
inputs_embeds = torch.cat(
|
170 |
+
(perturbed_inputs_embeds_1, perturbed_inputs_embeds_2),
|
171 |
+
dim=0,
|
172 |
+
)
|
173 |
+
|
174 |
+
kl_loss = -0.5 * (1 + log_var - log_var.exp()).mean(dim=-1).sum(dim=-1).mean()
|
175 |
+
contrastive_loss = (1 - F.cosine_similarity(perturbed_inputs_embeds_1, perturbed_inputs_embeds_2, dim=-1)).sum(dim=-1).mean()
|
176 |
+
emb_loss = kl_loss + contrastive_loss
|
177 |
+
|
178 |
+
if kl_loss.device == torch.device("cuda:0"):
|
179 |
+
self.log["kl_loss"] = kl_loss.item()
|
180 |
+
self.log["std"] = torch.exp(0.5 * log_var).mean().item()
|
181 |
+
self.log["contrastive_loss"] = contrastive_loss.item()
|
182 |
+
|
183 |
+
pass
|
184 |
+
elif do_task == "finetune":
|
185 |
+
inputs_embeds = inputs_embeds.detach()
|
186 |
+
inputs_embeds_refer = inputs_embeds.clone().detach()
|
187 |
+
shift_text_hidden_states = self.aligner_MLP(shift_text_hidden_states)
|
188 |
+
emb_origin_mask = text_labels != -100 # get output text pos
|
189 |
+
emb_shift_mask = shift_text_labels != -100
|
190 |
+
|
191 |
+
origin_padding_length = labels.shape[-1] - emb_origin_mask.shape[-1]
|
192 |
+
shift_padding_length = labels.shape[-1] - emb_shift_mask.shape[-1]
|
193 |
+
|
194 |
+
extend_emb_origin_mask = padding_tensor(emb_origin_mask, origin_padding_length, 1, False)
|
195 |
+
extend_emb_shift_mask = padding_tensor(emb_shift_mask, shift_padding_length, 1, False)
|
196 |
+
extend_shift_text_hidden_states = padding_tensor(shift_text_hidden_states, shift_padding_length, 1, 1e-9)
|
197 |
+
# check
|
198 |
+
extend_text_labels = padding_tensor(text_labels, origin_padding_length, 1, -100)
|
199 |
+
extend_shift_text_labels = padding_tensor(shift_text_labels, shift_padding_length, 1, -100)
|
200 |
+
|
201 |
+
assert torch.equal(
|
202 |
+
extend_text_labels[extend_emb_origin_mask],
|
203 |
+
extend_shift_text_labels[extend_emb_shift_mask]
|
204 |
+
), "{}\n{}\n{}\n{}".format(labels, extend_emb_origin_mask, extend_shift_text_labels, extend_emb_shift_mask)
|
205 |
+
|
206 |
+
inputs_embeds[extend_emb_origin_mask.unsqueeze(-1).expand_as(inputs_embeds)] = \
|
207 |
+
extend_shift_text_hidden_states[extend_emb_shift_mask.unsqueeze(-1).expand_as(extend_shift_text_hidden_states)].to(dtype=inputs_embeds.dtype)
|
208 |
+
|
209 |
+
if self.training:
|
210 |
+
contrastive_loss = (1 - F.cosine_similarity(inputs_embeds, inputs_embeds_refer, dim=-1)).sum(-1).mean()
|
211 |
+
emb_loss = contrastive_loss
|
212 |
+
if emb_loss.device == torch.device("cuda:0"):
|
213 |
+
self.log["contrastive_loss"] = contrastive_loss.item()
|
214 |
+
pass
|
215 |
+
elif do_task == "finetune_kd" :
|
216 |
+
#inputs_embeds = inputs_embeds.detach()
|
217 |
+
#inputs_embeds_refer = inputs_embeds.clone().detach()
|
218 |
+
#print(text_labels)
|
219 |
+
#print(sub_lengths.sum())
|
220 |
+
emb_origin_mask = text_labels != -100
|
221 |
+
|
222 |
+
fetch_lables_list = []
|
223 |
+
for batch in range(emb_origin_mask.shape[0]):
|
224 |
+
fetch_lables_list.append(text_labels[batch][emb_origin_mask[batch]])
|
225 |
+
shift_text_hidden_states = self.aligner_MLP(shift_text_hidden_states)
|
226 |
+
|
227 |
+
#split the shift_text_hidden_states
|
228 |
+
#[128006, 128000, 78191, 128007, 128000, 198, 128000]
|
229 |
+
maxn_length = sub_lengths.max() + 8
|
230 |
+
pad_ids = torch.full(size=(sub_lengths.shape[0], sub_lengths.shape[1], maxn_length), fill_value=self.pad_token_id, dtype=torch.long).to(shift_text_hidden_states.device)
|
231 |
+
|
232 |
+
pad_text_ids = torch.full(size=(sub_lengths.shape[0], sub_lengths.shape[1], maxn_length), fill_value=self.pad_token_id, dtype=torch.long).to(shift_text_hidden_states.device)
|
233 |
+
|
234 |
+
atten_mask = pad_ids.ne(self.pad_token_id)
|
235 |
+
#target_mask_part1 = pad_ids.ne(self.pad_token_id)
|
236 |
+
shift_text_hidden_states_slice = F.embedding(pad_ids, embed_tokens_weight, padding_idx=self.pad_token_id)
|
237 |
+
|
238 |
+
#print(shift_text_hidden_states_slice.shape,shift_text_hidden_states.shape)
|
239 |
+
for batch in range(sub_lengths.shape[0]):
|
240 |
+
cot=0
|
241 |
+
start_index = text_start_index[batch]
|
242 |
+
for index, sub_length in enumerate(sub_lengths[batch]):
|
243 |
+
if sub_length==-1:
|
244 |
+
break
|
245 |
+
#print(shift_text_hidden_states_slice[batch][index][:sub_length].shape, shift_text_hidden_states[batch][cot:cot+sub_length].shape)
|
246 |
+
eos_id = torch.IntTensor([128009]).to(inputs_embeds.device)
|
247 |
+
eos = self.model.embed_tokens(eos_id)
|
248 |
+
if index == 0:
|
249 |
+
text_prefix_ids = torch.IntTensor([128006, 128000, 65576, 128007, 128000, 198]).to(inputs_embeds.device)
|
250 |
+
preifx_embed = self.model.embed_tokens(text_prefix_ids)
|
251 |
+
pad_text_ids[batch][index][:sub_length+7] = torch.cat([text_prefix_ids, fetch_lables_list[batch][cot:cot+sub_length], eos_id],dim=0)
|
252 |
+
atten_mask[batch][index][:sub_length+7]=True
|
253 |
+
else:
|
254 |
+
text_prefix_ids = torch.IntTensor([128006, 128000, 65576, 128007, 128000, 198, 12800]).to(inputs_embeds.device)
|
255 |
+
preifx_embed = self.model.embed_tokens(text_prefix_ids)
|
256 |
+
pad_text_ids[batch][index][:sub_length+8] = torch.cat([text_prefix_ids, fetch_lables_list[batch][cot:cot+sub_length], eos_id], dim=0)
|
257 |
+
atten_mask[batch][index][:sub_length+8]=True
|
258 |
+
|
259 |
+
new_shift_text_hidden_states = torch.cat([preifx_embed, shift_text_hidden_states[batch][start_index+cot:start_index+cot+sub_length], eos], dim = 0)
|
260 |
+
shift_text_hidden_states_slice[batch][index][:new_shift_text_hidden_states.shape[0]] = new_shift_text_hidden_states
|
261 |
+
|
262 |
+
cot+=sub_length
|
263 |
+
shift_text_hidden_states_slice = shift_text_hidden_states_slice.reshape(shift_text_hidden_states_slice.shape[0]*shift_text_hidden_states_slice.shape[1],shift_text_hidden_states_slice.shape[2],shift_text_hidden_states_slice.shape[3])
|
264 |
+
|
265 |
+
|
266 |
+
padding_unit_targets = unit_targets.clone()
|
267 |
+
padding_unit_targets = torch.where(padding_unit_targets == IGNORE_TOKEN_ID, self.pad_token_id, padding_unit_targets)
|
268 |
+
target_mask_part = padding_unit_targets.ne(self.pad_token_id)
|
269 |
+
atten_mask = torch.cat([atten_mask, target_mask_part], dim = -1)
|
270 |
+
atten_mask = atten_mask.reshape(atten_mask.shape[0]*atten_mask.shape[1],atten_mask.shape[2])
|
271 |
+
|
272 |
+
pad_text_ids = pad_text_ids.reshape(pad_text_ids.shape[0]*pad_text_ids.shape[1],pad_text_ids.shape[2])
|
273 |
+
shift_text_embeddings = F.embedding(pad_text_ids, embed_tokens_weight, padding_idx=self.pad_token_id)
|
274 |
+
|
275 |
+
unit_target_slice = F.embedding(padding_unit_targets, embed_tokens_weight, padding_idx=self.pad_token_id)
|
276 |
+
# unit_target_slice = F.embedding(unit_targets, embed_tokens_weight, padding_idx=self.pad_token_id)
|
277 |
+
unit_target_slice = unit_target_slice.reshape(unit_target_slice.shape[0]*unit_target_slice.shape[1],unit_target_slice.shape[2],unit_target_slice.shape[3])
|
278 |
+
|
279 |
+
inputs_embeds = torch.cat([shift_text_hidden_states_slice, unit_target_slice], dim = 1)
|
280 |
+
|
281 |
+
ignore_ids = torch.full(size=(sub_lengths.shape[0], sub_lengths.shape[1], maxn_length), fill_value=IGNORE_TOKEN_ID, dtype=torch.long).to(shift_text_hidden_states.device)
|
282 |
+
unit_targets = torch.cat([ignore_ids,unit_targets],dim=-1)
|
283 |
+
unit_targets = unit_targets.reshape(unit_targets.shape[0]*unit_targets.shape[1],unit_targets.shape[2])
|
284 |
+
|
285 |
+
if self.training:
|
286 |
+
#print(shift_text_hidden_states_slice.shape, shift_text_embeddings.shape)
|
287 |
+
contrastive_loss = (1 - F.cosine_similarity(shift_text_hidden_states_slice, shift_text_embeddings, dim=-1)).sum(-1).mean()
|
288 |
+
emb_loss = contrastive_loss
|
289 |
+
if emb_loss.device == torch.device("cuda:0"):
|
290 |
+
self.log["contrastive_loss"] = contrastive_loss.item()
|
291 |
+
|
292 |
+
elif do_task == "finetune_kd_online":
|
293 |
+
shift_text_hidden_states = self.aligner_MLP(shift_text_hidden_states)
|
294 |
+
gold_inputs_embeds = inputs_embeds.clone()
|
295 |
+
for batch in range (inputs_embeds.shape[0]):
|
296 |
+
start_index = text_start_index[batch]
|
297 |
+
for slice_index in range (inputs_embeds.shape[1]):
|
298 |
+
sub_length= sub_lengths[batch][slice_index]
|
299 |
+
inputs_embeds[batch][slice_index][7:7+sub_length] = shift_text_hidden_states[batch][start_index+1:start_index+1+sub_length]
|
300 |
+
start_index += sub_length
|
301 |
+
if self.training:
|
302 |
+
#print(shift_text_hidden_states_slice.shape, shift_text_embeddings.shape)
|
303 |
+
contrastive_loss = ((1 - F.cosine_similarity(inputs_embeds, gold_inputs_embeds, dim=-1)) * attention_mask).sum(-1).mean()
|
304 |
+
emb_loss = contrastive_loss
|
305 |
+
if emb_loss.device == torch.device("cuda:0"):
|
306 |
+
self.log["contrastive_loss"] = contrastive_loss.item()
|
307 |
+
unit_embeds = F.embedding(unit_targets, embed_tokens_weight, padding_idx=self.pad_token_id)
|
308 |
+
|
309 |
+
inputs_embeds = torch.cat([inputs_embeds,unit_embeds], dim=2)
|
310 |
+
else:
|
311 |
+
inputs_embeds = self.aligner_MLP(inputs_embeds)
|
312 |
+
#[start_header_id] + _speaker + [end_header_id] + nl_tokens only for batch one!
|
313 |
+
units_ids = torch.IntTensor([[128009, 128006, 128000, 65576, 128007, 128000, 198]]).to(inputs_embeds.device)
|
314 |
+
units_prefix = self.model.embed_tokens(units_ids)
|
315 |
+
text_ids = torch.IntTensor([[128006, 128000, 65576, 128007, 128000, 198, 12800]]).to(inputs_embeds.device)
|
316 |
+
text_prefix = self.model.embed_tokens(text_ids)
|
317 |
+
inputs_embeds = torch.cat([text_prefix, inputs_embeds, units_prefix], dim = 1)
|
318 |
+
|
319 |
+
inputs_embeds = self.adapter(inputs_embeds)
|
320 |
+
if do_task == "finetune_kd":
|
321 |
+
return (emb_loss, inputs_embeds, unit_targets, atten_mask,)
|
322 |
+
else:
|
323 |
+
return (emb_loss, inputs_embeds)
|
324 |
+
|
325 |
+
def forward(
|
326 |
+
self,
|
327 |
+
input_ids: torch.LongTensor = None,
|
328 |
+
attention_mask: Optional[torch.Tensor] = None,
|
329 |
+
position_ids: Optional[torch.LongTensor] = None,
|
330 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
331 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
332 |
+
labels: Optional[torch.LongTensor] = None,
|
333 |
+
use_cache: Optional[bool] = None,
|
334 |
+
output_attentions: Optional[bool] = None,
|
335 |
+
output_hidden_states: Optional[bool] = None,
|
336 |
+
return_dict: Optional[bool] = None,
|
337 |
+
cache_position: Optional[torch.LongTensor] = None,
|
338 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
339 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
340 |
+
output_hidden_states = (
|
341 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
342 |
+
)
|
343 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
344 |
+
|
345 |
+
if inputs_embeds == None:
|
346 |
+
# inputs_embeds = self.model.embed_tokens(input_ids)
|
347 |
+
# share embedding 0501 by kkq
|
348 |
+
embed_tokens_weight = torch.cat(
|
349 |
+
[
|
350 |
+
self.model.embed_tokens.weight.detach(), self.unit_embedding.weight
|
351 |
+
],
|
352 |
+
dim = 0,
|
353 |
+
)
|
354 |
+
# print(embed_tokens_weight, embed_tokens_weight.shape)
|
355 |
+
inputs_embeds = F.embedding(input_ids, embed_tokens_weight, padding_idx=self.pad_token_id)
|
356 |
+
inputs_embeds = self.adapter(inputs_embeds)
|
357 |
+
|
358 |
+
outputs = self.model(
|
359 |
+
input_ids=None,
|
360 |
+
attention_mask=attention_mask,
|
361 |
+
past_key_values=past_key_values,
|
362 |
+
inputs_embeds=inputs_embeds,
|
363 |
+
use_cache=use_cache,
|
364 |
+
output_attentions=output_attentions,
|
365 |
+
output_hidden_states=output_hidden_states,
|
366 |
+
return_dict=return_dict,
|
367 |
+
)
|
368 |
+
hidden_states = outputs[0]
|
369 |
+
logits = self.lm_head(hidden_states)
|
370 |
+
|
371 |
+
loss = None
|
372 |
+
cr_loss = None
|
373 |
+
if labels != None:
|
374 |
+
shift_labels = labels
|
375 |
+
|
376 |
+
# Shift so that tokens < n predict n
|
377 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
378 |
+
shift_labels = shift_labels[..., 1:].contiguous()
|
379 |
+
|
380 |
+
loss_fct = CrossEntropyLoss()
|
381 |
+
|
382 |
+
shift_logits = shift_logits.view(-1, (self.config.vocab_size + self.config.unit_vocab_size))
|
383 |
+
shift_labels = shift_labels.view(-1)
|
384 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
385 |
+
|
386 |
+
loss = loss_fct(shift_logits, shift_labels)
|
387 |
+
|
388 |
+
if loss.device == torch.device("cuda:0"):
|
389 |
+
self.log["unit_loss"] = loss.item()
|
390 |
+
|
391 |
+
if cr_loss != None:
|
392 |
+
target_scale = loss.item() * 0.2
|
393 |
+
cr_loss_weight = target_scale / cr_loss.item() if cr_loss > target_scale else 1.0
|
394 |
+
loss = loss + cr_loss_weight * cr_loss
|
395 |
+
|
396 |
+
if loss.device == torch.device("cuda:0") and (self.current_step - 10) % 100 == 0:
|
397 |
+
print(self.log, loss.device)
|
398 |
+
|
399 |
+
if not return_dict:
|
400 |
+
output = (logits,) + outputs[1:]
|
401 |
+
return (loss,) + output if loss is not None else output
|
402 |
+
|
403 |
+
return CausalLMOutputWithPast(
|
404 |
+
loss=loss,
|
405 |
+
logits=logits,
|
406 |
+
past_key_values=outputs.past_key_values,
|
407 |
+
hidden_states=outputs.hidden_states,
|
408 |
+
attentions=outputs.attentions,
|
409 |
+
)
|
410 |
+
|
411 |
+
AutoConfig.register("T2ULlama", T2ULlamaConfig)
|
412 |
+
AutoModelForCausalLM.register(T2ULlamaConfig, T2ULlamaForCausalLM)
|
show_case/1.wav
ADDED
Binary file (72.1 kB). View file
|
|
show_case/2.wav
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:03198985b12bd892d05b8ae9b2e6c8303b15b0a570eea9647b2e314332340711
|
3 |
+
size 165742
|
show_case/Translate_de_audio_prompt.wav
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:aee2ba64475fb597ec0a68da4e1a552fc4463bca0e9393d9f170e37f563d7792
|
3 |
+
size 288164
|
text_to_speech.py
ADDED
@@ -0,0 +1,326 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fairseq.dataclass.configs import FairseqConfig
|
2 |
+
from fairseq import utils
|
3 |
+
from fairseq.models.text_to_speech.vocoder import CodeHiFiGANVocoder
|
4 |
+
from fairseq import checkpoint_utils, options, tasks, utils
|
5 |
+
from fairseq.distributed import utils as distributed_utils
|
6 |
+
import torch
|
7 |
+
import json
|
8 |
+
from tqdm import tqdm
|
9 |
+
import random
|
10 |
+
import soundfile as sf
|
11 |
+
import numpy as np
|
12 |
+
import ast
|
13 |
+
import time
|
14 |
+
import math
|
15 |
+
from fairseq.dataclass.utils import convert_namespace_to_omegaconf
|
16 |
+
from fairseq.token_generation_constraints import pack_constraints, unpack_constraints
|
17 |
+
from fairseq_cli.generate import get_symbols_to_strip_from_output
|
18 |
+
from collections import namedtuple
|
19 |
+
import sys
|
20 |
+
from argparse import Namespace
|
21 |
+
import argparse
|
22 |
+
import sentencepiece as spm
|
23 |
+
import re
|
24 |
+
|
25 |
+
Batch = namedtuple("Batch", "ids src_tokens src_lengths constraints")
|
26 |
+
Translation = namedtuple("Translation", "src_str hypos pos_scores alignments")
|
27 |
+
|
28 |
+
def make_batches(lines, cfg, task, max_positions, encode_fn):
|
29 |
+
def encode_fn_target(x):
|
30 |
+
return encode_fn(x)
|
31 |
+
|
32 |
+
if cfg.generation.constraints:
|
33 |
+
# Strip (tab-delimited) contraints, if present, from input lines,
|
34 |
+
# store them in batch_constraints
|
35 |
+
batch_constraints = [list() for _ in lines]
|
36 |
+
for i, line in enumerate(lines):
|
37 |
+
if "\t" in line:
|
38 |
+
lines[i], *batch_constraints[i] = line.split("\t")
|
39 |
+
|
40 |
+
# Convert each List[str] to List[Tensor]
|
41 |
+
for i, constraint_list in enumerate(batch_constraints):
|
42 |
+
batch_constraints[i] = [
|
43 |
+
task.target_dictionary.encode_line(
|
44 |
+
encode_fn_target(constraint),
|
45 |
+
append_eos=False,
|
46 |
+
add_if_not_exist=False,
|
47 |
+
)
|
48 |
+
for constraint in constraint_list
|
49 |
+
]
|
50 |
+
|
51 |
+
if cfg.generation.constraints:
|
52 |
+
constraints_tensor = pack_constraints(batch_constraints)
|
53 |
+
else:
|
54 |
+
constraints_tensor = None
|
55 |
+
|
56 |
+
tokens, lengths = task.get_interactive_tokens_and_lengths(lines, encode_fn)
|
57 |
+
|
58 |
+
itr = task.get_batch_iterator(
|
59 |
+
dataset=task.build_dataset_for_inference(
|
60 |
+
tokens, lengths, constraints=constraints_tensor
|
61 |
+
),
|
62 |
+
max_tokens=cfg.dataset.max_tokens,
|
63 |
+
max_sentences=cfg.dataset.batch_size,
|
64 |
+
max_positions=max_positions,
|
65 |
+
ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test,
|
66 |
+
).next_epoch_itr(shuffle=False)
|
67 |
+
for batch in itr:
|
68 |
+
ids = batch["id"]
|
69 |
+
src_tokens = batch["net_input"]["src_tokens"]
|
70 |
+
src_lengths = batch["net_input"]["src_lengths"]
|
71 |
+
constraints = batch.get("constraints", None)
|
72 |
+
|
73 |
+
yield Batch(
|
74 |
+
ids=ids,
|
75 |
+
src_tokens=src_tokens,
|
76 |
+
src_lengths=src_lengths,
|
77 |
+
constraints=constraints,
|
78 |
+
)
|
79 |
+
|
80 |
+
def tokenize(inputs, sp):
|
81 |
+
text = re.sub(r'[^\w\s]', '', inputs.lower())
|
82 |
+
inputs = ' '.join(sp.EncodeAsPieces(text))
|
83 |
+
# print(inputs)
|
84 |
+
return inputs
|
85 |
+
|
86 |
+
def get_t2u_config(model, beam=5):
|
87 |
+
|
88 |
+
sys.argv = [
|
89 |
+
"fairseq-interactive",
|
90 |
+
"libri_t2u",
|
91 |
+
"--path", model,
|
92 |
+
"--gen-subset", "valid",
|
93 |
+
"--max-len-b", "1024",
|
94 |
+
"--max-source-positions", "500",
|
95 |
+
"--max-target-positions", "1024",
|
96 |
+
"--beam", str(beam),
|
97 |
+
"--results-path", "decode"
|
98 |
+
]
|
99 |
+
|
100 |
+
parser = options.get_interactive_generation_parser()
|
101 |
+
args = options.parse_args_and_arch(parser)
|
102 |
+
# distributed_utils.call_main(convert_namespace_to_omegaconf(args), load_text2units_model)
|
103 |
+
return convert_namespace_to_omegaconf(args)
|
104 |
+
|
105 |
+
def load_text2units_model(cfg: FairseqConfig, device):
|
106 |
+
|
107 |
+
if isinstance(cfg, Namespace):
|
108 |
+
cfg = convert_namespace_to_omegaconf(cfg)
|
109 |
+
|
110 |
+
utils.import_user_module(cfg.common)
|
111 |
+
if cfg.interactive.buffer_size < 1:
|
112 |
+
cfg.interactive.buffer_size = 1
|
113 |
+
if cfg.dataset.max_tokens is None and cfg.dataset.batch_size is None:
|
114 |
+
cfg.dataset.batch_size = 1
|
115 |
+
|
116 |
+
assert (
|
117 |
+
not cfg.generation.sampling or cfg.generation.nbest == cfg.generation.beam
|
118 |
+
), "--sampling requires --nbest to be equal to --beam"
|
119 |
+
assert (
|
120 |
+
not cfg.dataset.batch_size
|
121 |
+
or cfg.dataset.batch_size <= cfg.interactive.buffer_size
|
122 |
+
), "--batch-size cannot be larger than --buffer-size"
|
123 |
+
|
124 |
+
# Fix seed for stochastic decoding
|
125 |
+
if cfg.common.seed is not None and not cfg.generation.no_seed_provided:
|
126 |
+
np.random.seed(cfg.common.seed)
|
127 |
+
utils.set_torch_seed(cfg.common.seed)
|
128 |
+
|
129 |
+
use_cuda = torch.cuda.is_available() and not cfg.common.cpu
|
130 |
+
|
131 |
+
# Setup task, e.g., translation
|
132 |
+
task = tasks.setup_task(cfg.task)
|
133 |
+
|
134 |
+
# Load ensemble
|
135 |
+
overrides = ast.literal_eval(cfg.common_eval.model_overrides)
|
136 |
+
models, _model_args = checkpoint_utils.load_model_ensemble(
|
137 |
+
utils.split_paths(cfg.common_eval.path),
|
138 |
+
arg_overrides=overrides,
|
139 |
+
task=task,
|
140 |
+
suffix=cfg.checkpoint.checkpoint_suffix,
|
141 |
+
strict=(cfg.checkpoint.checkpoint_shard_count == 1),
|
142 |
+
num_shards=cfg.checkpoint.checkpoint_shard_count,
|
143 |
+
)
|
144 |
+
|
145 |
+
# Set dictionaries
|
146 |
+
src_dict = task.source_dictionary
|
147 |
+
tgt_dict = task.target_dictionary
|
148 |
+
|
149 |
+
# Optimize ensemble for generation
|
150 |
+
for model in models:
|
151 |
+
if model is None:
|
152 |
+
continue
|
153 |
+
if cfg.common.fp16:
|
154 |
+
model.half()
|
155 |
+
if use_cuda and not cfg.distributed_training.pipeline_model_parallel:
|
156 |
+
model.cuda()
|
157 |
+
model.prepare_for_inference_(cfg)
|
158 |
+
|
159 |
+
# Initialize generator
|
160 |
+
generator = task.build_generator(models, cfg.generation)
|
161 |
+
|
162 |
+
# Handle tokenization and BPE
|
163 |
+
tokenizer = task.build_tokenizer(cfg.tokenizer)
|
164 |
+
bpe = task.build_bpe(cfg.bpe)
|
165 |
+
|
166 |
+
return {
|
167 |
+
"models": models,
|
168 |
+
"generator": generator,
|
169 |
+
"tokenizer": tokenizer,
|
170 |
+
"bpe": bpe,
|
171 |
+
"task": task,
|
172 |
+
"src_dict": src_dict,
|
173 |
+
"tgt_dict": tgt_dict,
|
174 |
+
"use_cuda": use_cuda
|
175 |
+
}
|
176 |
+
|
177 |
+
def gen_units(model, cfg, inputs):
|
178 |
+
inputs = [inputs]
|
179 |
+
|
180 |
+
models = model['models']
|
181 |
+
generator = model['generator']
|
182 |
+
tokenizer = model['tokenizer']
|
183 |
+
bpe = model['bpe']
|
184 |
+
task = model['task']
|
185 |
+
src_dict = model['src_dict']
|
186 |
+
tgt_dict = model['tgt_dict']
|
187 |
+
use_cuda = model['use_cuda']
|
188 |
+
|
189 |
+
def encode_fn(x):
|
190 |
+
if tokenizer is not None:
|
191 |
+
x = tokenizer.encode(x)
|
192 |
+
if bpe is not None:
|
193 |
+
x = bpe.encode(x)
|
194 |
+
return x
|
195 |
+
|
196 |
+
def decode_fn(x):
|
197 |
+
if bpe is not None:
|
198 |
+
x = bpe.decode(x)
|
199 |
+
if tokenizer is not None:
|
200 |
+
x = tokenizer.decode(x)
|
201 |
+
return x
|
202 |
+
|
203 |
+
align_dict = utils.load_align_dict(cfg.generation.replace_unk)
|
204 |
+
|
205 |
+
max_positions = utils.resolve_max_positions(
|
206 |
+
task.max_positions(), *[model.max_positions() for model in models]
|
207 |
+
)
|
208 |
+
|
209 |
+
start_id = 0
|
210 |
+
results = []
|
211 |
+
for batch in make_batches(inputs, cfg, task, max_positions, encode_fn):
|
212 |
+
print("[INFO_DEBUG]", batch)
|
213 |
+
bsz = batch.src_tokens.size(0)
|
214 |
+
src_tokens = batch.src_tokens
|
215 |
+
src_lengths = batch.src_lengths
|
216 |
+
constraints = batch.constraints
|
217 |
+
if use_cuda:
|
218 |
+
src_tokens = src_tokens.cuda()
|
219 |
+
src_lengths = src_lengths.cuda()
|
220 |
+
if constraints is not None:
|
221 |
+
constraints = constraints.cuda()
|
222 |
+
|
223 |
+
sample = {
|
224 |
+
"net_input": {
|
225 |
+
"src_tokens": src_tokens,
|
226 |
+
"src_lengths": src_lengths,
|
227 |
+
},
|
228 |
+
}
|
229 |
+
translate_start_time = time.time()
|
230 |
+
translations = task.inference_step(
|
231 |
+
generator, models, sample, constraints=constraints
|
232 |
+
)
|
233 |
+
translate_time = time.time() - translate_start_time
|
234 |
+
list_constraints = [[] for _ in range(bsz)]
|
235 |
+
if cfg.generation.constraints:
|
236 |
+
list_constraints = [unpack_constraints(c) for c in constraints]
|
237 |
+
for i, (id, hypos) in enumerate(zip(batch.ids.tolist(), translations)):
|
238 |
+
src_tokens_i = utils.strip_pad(src_tokens[i], tgt_dict.pad())
|
239 |
+
constraints = list_constraints[i]
|
240 |
+
results.append(
|
241 |
+
(
|
242 |
+
start_id + id,
|
243 |
+
src_tokens_i,
|
244 |
+
hypos,
|
245 |
+
{
|
246 |
+
"constraints": constraints,
|
247 |
+
"time": translate_time / len(translations),
|
248 |
+
},
|
249 |
+
)
|
250 |
+
)
|
251 |
+
|
252 |
+
# print(results)
|
253 |
+
|
254 |
+
units = []
|
255 |
+
for id_, _, hypos, info in sorted(results, key=lambda x: x[0]):
|
256 |
+
print("W-{}\t{:.3f}\tseconds".format(id_, info["time"]))
|
257 |
+
|
258 |
+
# Process top predictions
|
259 |
+
for hypo in hypos[: min(len(hypos), cfg.generation.nbest)]:
|
260 |
+
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
|
261 |
+
hypo_tokens=hypo["tokens"].int().cpu(),
|
262 |
+
src_str="",
|
263 |
+
alignment=hypo["alignment"],
|
264 |
+
align_dict=align_dict,
|
265 |
+
tgt_dict=tgt_dict,
|
266 |
+
remove_bpe=cfg.common_eval.post_process,
|
267 |
+
extra_symbols_to_ignore=get_symbols_to_strip_from_output(generator),
|
268 |
+
)
|
269 |
+
|
270 |
+
units.append(list(map(int, hypo_str.split(' '))))
|
271 |
+
|
272 |
+
return units
|
273 |
+
|
274 |
+
def get_vocoder_config(vocoder, config):
|
275 |
+
|
276 |
+
args = argparse.Namespace(
|
277 |
+
vocoder=vocoder,
|
278 |
+
vocoder_cfg=config,
|
279 |
+
dur_prediction=True,
|
280 |
+
speaker_id=1,
|
281 |
+
cpu=False
|
282 |
+
)
|
283 |
+
|
284 |
+
return args
|
285 |
+
|
286 |
+
def load_units_vocoder(args, device):
|
287 |
+
with open(args.vocoder_cfg) as f:
|
288 |
+
vocoder_cfg = json.load(f)
|
289 |
+
vocoder = CodeHiFiGANVocoder(args.vocoder, vocoder_cfg).to(device)
|
290 |
+
|
291 |
+
multispkr = vocoder.model.multispkr
|
292 |
+
if multispkr:
|
293 |
+
num_speakers = vocoder_cfg.get(
|
294 |
+
"num_speakers", 200
|
295 |
+
) # following the default in codehifigan to set to 200
|
296 |
+
assert (
|
297 |
+
args.speaker_id < num_speakers
|
298 |
+
), f"invalid --speaker-id ({args.speaker_id}) with total #speakers = {num_speakers}"
|
299 |
+
|
300 |
+
return vocoder, num_speakers if multispkr else 1, 'cuda' in device
|
301 |
+
|
302 |
+
def gen_wav(vocoder, args, data, device):
|
303 |
+
vocoder, num_speakers, use_cuda = vocoder
|
304 |
+
res = []
|
305 |
+
for i, d in enumerate(data): # tqdm is removed for cleaner streaming
|
306 |
+
x = {
|
307 |
+
"code": torch.LongTensor(d).view(1, -1).to(device),
|
308 |
+
}
|
309 |
+
suffix = ""
|
310 |
+
|
311 |
+
multispkr = vocoder.model.multispkr
|
312 |
+
if multispkr:
|
313 |
+
spk = (
|
314 |
+
random.randint(0, num_speakers - 1)
|
315 |
+
if args.speaker_id == -1
|
316 |
+
else args.speaker_id
|
317 |
+
)
|
318 |
+
suffix = f"_spk{spk}"
|
319 |
+
x["spkr"] = torch.LongTensor([spk]).view(1, 1)
|
320 |
+
|
321 |
+
x = utils.move_to_cuda(x) if use_cuda else x
|
322 |
+
wav = vocoder(x, args.dur_prediction).detach().cpu().numpy()
|
323 |
+
|
324 |
+
res.append(wav)
|
325 |
+
|
326 |
+
return res[0]
|